sampling activity: display hotspots % including subroutines

[kabelgleichung]

Hello!

I am using VTune's sampling activity on Linux/VTune 9.1 build 226, and was wondering if it is possible to display the hotspot-functions CPU_CLK_UNHALTED.CORE % as accumulated values with their subroutines (like OProfile does with it's callgraph sampling ability).
I am viewing the hotspot's results as follows:
vtl view -ar a1 -hf -mn libavcodec.so,g722_audio_pwplugin.so

For some reasons (slowdown of the application), I cannot use VTune's callgraph activity.

Thanks a lot!

[Vladimir_T_Intel]

Quoting - kabelgleichung

Hello!

I am using VTune's sampling activity on Linux/VTune 9.1 build 226, and was wondering if it is possible to display the hotspot-functions CPU_CLK_UNHALTED.CORE % as accumulated values with their subroutines (like OProfile does with it's callgraph sampling ability).
I am viewing the hotspot's results as follows:
vtl view -ar a1 -hf -mn libavcodec.so,g722_audio_pwplugin.so

For some reasons (slowdown of the application), I cannot use VTune's callgraph activity.

Thanks a lot!

This is the general limitation of the sampling technology - you get the samples that hit the concrete function without any connection to children or parent functions, i.e. flat profile.

However, you might want to try Intel Performance Tuning Utility, which provides Statistical Call Graph technology and may help to find what you need.

[kabelgleichung]

Quoting - Vladimir Tsymbal (Intel)

This is the general limitation of the sampling technology - you get the samples that hit the concrete function without any connection to children or parent functions, i.e. flat profile.

However, you might want to try Intel Performance Tuning Utility, which provides Statistical Call Graph technology and may help to find what you need.

Thank you for pointing me to the Performance Tuning Utility. I tried it and was able to create sampled call graph data. However, the libraries I'm trying to profile consist of some pretty short (in instructions) functions, that are not covered by the smallest configureable sampling interval of 1ms (when using oprofile with it's smallest sampling frequence of 100000 clock cycles I get an interval size of 0.0625ms with my 1,6GHz CPU).

Can you give me any hints to create sampled-call graphs with smaller intervals using VTune or the Performance Tuning Utility?

Thanks a lot again!

[Vladimir_T_Intel]

Quoting - kabelgleichung

Thank you for pointing me to the Performance Tuning Utility. I tried it and was able to create sampled call graph data. However, the libraries I'm trying to profile consist of some pretty short (in instructions) functions, that are not covered by the smallest configureable sampling interval of 1ms (when using oprofile with it's smallest sampling frequence of 100000 clock cycles I get an interval size of 0.0625ms with my 1,6GHz CPU).

Can you give me any hints to create sampled-call graphs with smaller intervals using VTune or the Performance Tuning Utility?

Thanks a lot again!

Smaller intervals provides EBS only, which doesn't provide stacks, though. However, the general solution would be to transform the analyzed application by making those small functions called many times. This would statistically increase the probability of samples to hit the address range within the functions.